Published
Everyone swears scalping is where the money is. So I tested it — not with vibes, with data. Thousands of setups. Two years of candles. Real fees baked in, not the fantasy zero-fee version.
I wanted it to work. It didn't. And the way it failed is the part worth your five minutes.
High win rate, negative account
The most seductive trap is win rate. Plenty of configs hit 60–70% winners and looked brilliant. But when you added real fees and let the losers run their full size, the net result was negative. A strategy that wins small 65% of the time and loses big 35% of the time is a losing strategy with a flattering scoreboard.
Across the broad sweep, the clean crypto-scalp edges that survived honest, out-of-sample testing was essentially zero. The good-looking ones were propped up by averaging down or by extra-wide stops that quietly carried huge hidden drawdowns.
Fees are the silent killer
At a typical 0.055% taker fee per side, every round trip costs about 0.11%. Scalp 10 times a day and that is over 1% a day in fees alone — more than most real edges produce. The math is brutal: the more you trade, the more certain it is that fees, not the market, decide your result. (You can see this directly with the break-even calculator — price has to clear your entry plus both fees before you make a cent.)
What did survive
Almost nothing fast. The approaches that held up under walk-forward testing were slow and boring: trend-following on broad indices (be long while above the 200-day average, step aside below it) delivered better risk-adjusted returns than buy-and-hold with roughly half the drawdown over 18 years, across the 2008, 2020 and 2022 crashes. Not exciting. Just durable.
The takeaway
If a strategy needs you to trade constantly, be suspicious. The edge usually lives in doing less: fewer trades, lower fees, controlled risk, and patience. Three habits beat almost every clever scalping system we tested — size by risk with the position size calculator, know your liquidation price, and judge positions by realized P/L, not by a hopeful average.